AI raises the stakes for data in use

AI makes the regulatory relevance of confidential computing even clearer. AI workloads often process highly sensitive inputs: prompts, embeddings, documents, training data, fine-tuning datasets, model weights, and inference outputs. In some cases, the model itself may be valuable intellectual property. In others, the data being processed may be personal, confidential, classified, or commercially sensitive. The EU AI Act requires high-risk AI systems to achieve appropriate levels of accuracy, robustness and cybersecurity throughout their lifecycle.

That lifecycle framing is important. Security is not a one-time property of the model. It depends on how the model is deployed, what data it processes, how it is updated, how it is monitored, and how its surrounding infrastructure is protected. Confidential computing can help secure parts of that lifecycle, especially where organizations want to run AI workloads on shared, remote, or hybrid infrastructure without exposing sensitive data or models to the underlying platform. This is why confidential AI is becoming such an important use case.

It is not only about privacy. It is also about protecting intellectual property, reducing third-party risk, supporting regulated inference, and enabling organizations to adopt AI without giving up control over sensitive assets. The same idea is starting to appear in AI policy discussions beyond traditional compliance. For example, public comments related to the US American AI Exports Program have recommended accelerator-level confidential computing as one way to protect AI model weights and systems in export packages. That is not the same as a regulatory mandate, but it shows how confidential computing is entering wider discussions about AI security, trust, and geopolitical risk.